Text-to-Speech

Text-to-Speech, TTS, synthesizes spoken audio from written text for voice interfaces, audiobooks, accessibility, and voice cloning.

The usual pipeline analyzes text (abbreviations, numbers, punctuation), predicts prosody (intonation, stress, timing), then synthesizes a waveform. Early systems spliced recorded fragments, concatenative synthesis. Neural TTS sounds much closer to people. Tacotron and its successors generate mel spectrograms from text.

Vocoders such as WaveNet, WaveGlow, and HiFi-GAN turn those spectrograms into waveforms. VALL-E and Bark can clone a voice from seconds of reference audio, which is useful and also raises synthetic-media risk. Zero-shot TTS can hit new voices without fine-tuning.

Control of prosody, emphasis, emotion, and speaking rate is still hard. Multilingual TTS must handle several languages and code-switching. Streaming TTS with low latency is required for live assistants. Uses include accessibility, content production, assistants, navigation, language learning, and entertainment. Listeners notice bad rhythm, wrong pronunciations, and robotic pitch immediately.

" and "2026" into speakable words. Prosody prediction sets pitch and duration. Then a vocoder writes samples. Tacotron-style models predict mel spectrograms; WaveNet, WaveGlow, and HiFi-GAN invert them. Concatenative systems could not match that naturalness. VALL-E and Bark clone from a few seconds of audio, which is enough for accessibility voices and also for spoofing.

Streaming assistants cannot wait for a full paragraph; they need low-latency chunks. Listeners still catch bad stress and robotic intonation, so prosody control remains the hard part. Tacotron (2017) mapped text to spectrograms with a sequence model. Today's voice APIs still do text in, audio out.